You are on page 1of 11

Article

Transactions of the Institute of


Measurement and Control
Firefly artificial intelligence technique 2019, Vol. 41(10) 2875–2885
Ó The Author(s) 2019

for model order reduction with Article reuse guidelines:


sagepub.com/journals-permissions
DOI: 10.1177/0142331218814288
substructure preservation journals.sagepub.com/home/tim

Othman Alsmadi1, Adnan Al-Smadi2 and Esra’a Gharaibeh2

Abstract
Model order reduction (MOR) is a process of finding a lower order model for the original high order system with reasonable accuracy in order to sim-
plify analysis, design, modeling and simulation for large complex systems. It is desirable that the reduced order model preserves the fundamental prop-
erties of the original system. This paper presents a new MOR technique of multi-input multi-output systems utilizing the firefly algorithm (FA) as an
artificial intelligence technique. The reduction operation is proposed to maintain the exact dominant dynamics in the reduced order model with the
advantage of substructure preservation. This is mainly possible for systems that are characterized as multi-time scale systems. Obtaining the reduced
order model is achieved by minimizing the fitness function that is related to the error between the full and reduced order models’ responses. The new
approach is compared with recently published work on firefly optimization for MOR, in addition to three other artificial intelligence techniques; namely,
invasive weed optimization, particle swarm optimization and genetic algorithm. As a result, simulations show the potential of the FA for the process of
MOR.

Keywords
Model order reduction, firefly optimization, artificial intelligence, singular perturbation approximation, dominant dynamics

Introduction becomes necessary to reduce the computational complexity in


such cases. For MOR, there are many scenarios that can be
For the purpose of analysis and design, physical systems are applied. One produces reduced order models that are new
mathematically modeled. The main objective of such analysis and not related to the original models, in terms of their criti-
and design is for controlling the process and or enhancing its cal frequencies for either single-input single-output (SISO) or
performance (Alsmadi et al., 2011a, 2011b, 2016; Antoulas multi-input multi-output (MIMO) systems. Another scenario
et al., 2001; Yadav et el., 2012). For some practical systems, that achieves model reduction is the one that tries to keep the
mathematical modeling using higher order differential equa- original system’s important features regardless of its critical
tions yields complex large order multi-time scale systems. To frequencies. Due to meaningful physical interpretation in
reduce the complexity of such systems, it is desirable to obtaining similar models and due to minimum changes in the
obtain a fairly accurate lower order approximation of the original system, it was noted that the second scenario is more
original system while keeping the dominant behavior of the preferable, if possible (Alsmadi et al., 2011a, 2014).
original system. This reduction will result in reducing the For MOR, there are several optimization methods that
hardware complexity and thus simplifying the controller provide different solutions (Alsmadi et al., 2016; Chidambara,
design (Alsmadi et al., 2014a). 1967; Chen and Shieh, 1968; Davison, 1966; Mukherjee et al.,
Model order reduction (MOR) has played a significant 2005; Salimbahrami and Lohmann, 2006; Wilson, 1970;
role in control systems (Benner et al., 2015; Chen and Shieh, Yadav et al., 2012). One of the recently developed evolution-
1968; Davison, 1966; Gugercin and Antoulas, 2004). It ary techniques is the firefly algorithm (FA) optimization
focuses on the properties of dynamical systems in application method, which was developed by Yang (2009). This optimiza-
to reduce their complexity, while their input-output funda- tion method was described as a stochastic, nature-inspired,
mental properties of original system such as stability and meta-heuristic algorithm that can be used to solve difficult
dominant dynamics are preserved as much as possible. MOR
aims to simplify analysis, design, modeling and simulation of 1
such large complex systems. Moreover, the error between Electrical Engineering Department, The University of Jordan, Jordan
2
Electronics Engineering Department, Yarmouk University, Jordan
original and reduced order system is required to be as small
as possible. Corresponding author:
Due to complexity and large system dimension, most of Othman Alsmadi, Electrical Engineering Department, The University of
the modern numerical models of real-life processes face diffi- Jordan, Amman 11942, Jordan.
culties when used in numerical simulations. Hence, MOR Email: othman_mk@yahoo.com
2876 Transactions of the Institute of Measurement and Control 41(10)

optimization problems (Fister et al., 2013). Heuristic opera- The reduced order model, on the other hand, is obtained as
tion is referred to as discovering solutions by trial and error in
a reasonable amount of time and that there is no guarantee x_ r (t) = Ar xr (t) + Br u(t)
ð3Þ
that optimal solutions are reached. Stochastic is described as yr (t) = Cr xr (t)
it uses some kind of randomization in searching for a set of
solutions. The FA is inspired by the flashing behavior of fire- where Ar 2 <nr 3 nr , Br 2 <nr 3 m and Cr 2 <p 3 nr : Note that
flies as they search for their targets. In which it attracts other the feed-through Dr has been eliminated in order to produce
fireflies formatting. The idea of attractiveness and informa- a strictly proper model. The model is motivated by the SP
tion passing is what leads to the FA inspiration (Fister et al., technique. The SP objective in the reduced order model is
2013). achieved by maintaining the full order dominant dynamics in
In this paper, the ability of the FA technique to produce the reduced order model, as motivated by singular perturba-
relatively acceptable reduced order models while maintaining tion method. To achieve such an objective, the multi-time-
the system’s main characteristics will be investigated. The idea scale state matrix Ar is designed to have the following form
has been motivated by the singular perturbation approxima-
2 3
tion (SPA) (Alsmadi et al., 2014a). The paper proposes a new l1 a12 a13 a14  a1nr
MOR technique of MIMO systems utilizing the firefly opti- 60 l2 a23 a24  7
6 .. .. .. 7
mization method. The reduction process is performed in two 6 7
6 . . . 7
parts; the first one uses the dominant pole method, while the 6 .. 7
6 : 0 lb . 7
second utilizes the firefly optimization method. The reduction 6 7
Ar = 6
6 : 0 s1 a1 7
7
operation is proposed to maintain the exact dominant 6 : 7
6 0 a1 s1 7
dynamics in the reduced order model with the advantage of 6 .. 7
substructure preservation (SP). This is mainly possible for 6 0 0 . a(nr 2)nr 7
6 7
4 0 sp a 5
systems that are characterized as multi-time scale systems. p
Obtaining the reduced order model is achieved by minimizing 0 : : : 0 ap sp
the fitness function that is related to the error between the full ð4Þ
and reduced order models’ responses. Simulation results show
the potential of the FA as an artificial intelligence technique where the original system dominant eigenvalues (real and/or
for the process of MOR. The new technique is compared with complex) are preserved in the diagonal, with time scale
recently published work on firefly for MOR. The results pre- arrangement set as li, i = 1, 2, . b (real) and si 6ai , i = 1,
sented in this paper show the superiority of the proposed FA 2, . p (complex). Notice that for this reduced-order model
method over the other methods. In addition, the work pre- nr = (b + 2p)\n, and a reduction can be performed to any
sented will compare the results of the proposed FA reduction time scale as appropriate. It is important to mention that for
technique with some other artificial intelligence optimization an efficient MOR of multi-time-scale systems (i.e. dominant
techniques; namely, invasive weed optimization (IWO) and none dominant dynamic categories), the following condi-
(Mehrabian and Lucas, 2006), particle swarm optimization tion is to be satisfied (with li being negative real parts)
(PSO) (Kennedy and Eberhart, 1995) and genetic algorithm
(GA) (Abo-Hammour et al., 2011; Alsmadi et al., 2011c). ldominant dl1 . l2 .    . lb , l . s, and s1 . s2 .    . sp
The paper is organized as follows. In Section 2, we present ð5Þ
the mathematical framework for developing the MOR model
parameters estimation method. In Section 3, we present simu- where all of the elements in (5) are considered as dominant
lation results. In Section 4, a discussion is presented. Finally, dynamics (the real and complex distinct eigenvalues). The
we give some concluding remarks in Section 5. modal form is chosen, as given in (4), which implies that the
elements as aij (i and j = 1,.nr) are set to zero. On the other
hand, the reduction operation is performed by eliminating the
Problem formulation non-dominant dynamics given by

Background lnon dominant dlnr + 1 ,    , sn ð6Þ


th
Consider the following n order multi-time-scale linear time For proper reduced order modeling, the order of the reduced
invariant system model can be obtained as the difference between the value of
the full model order, n, and the number of non-dominant
x_ f (t) = Af xf (t) + Bf u(t) dynamics (lnon dominant dlnr + 1 ,    , sn ), to provide mini-
ð1Þ
yf (t) = Cf xf (t) + Df u(t) mum error modeling. The designer may go beyond this value;
however, a higher error would be expected.
where xf (t) 2 <n is the state vector, u(t) 2 <p and yf (t) 2 <m
are the input and output vectors, respectively. The matrices Af,
Bf, Cf, and Df represent the full order system matrices with their FA
appropriate dimensions. The system dynamics are given by FA is one of the recent swarm intelligence methods developed
by Yang (2009) and is a stochastic, nature-inspired, meta-
l = ½l1 l2    ln  ð2Þ heuristic algorithm that can be used to solve difficult
Alsmadi et al. 2877

optimization problems. Fireflies use a system of flashes to advantages suggest the use of the FA to increase the efficiency
communicate. They use their light to attract others. A firefly without deterioration of approximation accuracy for model
emits light from a tiny organ called a lantern, where a bio- reduction.
chemical reaction takes place. The reaction releases energy in After initialization, each firefly is compared against all of
the form of light. Every species of flashing firefly has its own the other fireflies, and will move towards every brighter firefly
pattern. These unique patterns let males and females of the encountered. Once a bright firefly is found, the distance
same species recognize one another in the dark (Fister et al., between the fireflies, has to be calculated. Different forms of
2013). distance calculation can be used; however, in general the
To govern the algorithm and create a modeled firefly’s Cartesian distance is appropriate. The Cartesian distance
behavior, there are three notes to be taken. Firstly, the fire- between two fireflies in D-dimensional space can be calcu-
flies are unisex; therefore, any firefly could be attracted to lated as follows (Alsmadi et al., 2016)
any of the other fireflies. Secondly, the attractiveness is deter- vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
mined by their brightness, where a less bright firefly will move u D
  uX
towards a brighter one. Finally, the brightness of a firefly is rij = xi  xj = t
  (xi, d  xj, d )2 ð9Þ
proportional to the value of the function being minimized d=1
(Fister et al., 2013).
The locations of the fireflies must be considered when com- where xi and xj are the position vectors for firefly i and j,
paring the brightness of any two fireflies. In the real world, if respectively, with xi (n) representing the position value for the
a firefly is searching for another, it can only see so far (Fister nth dimension. The movement of a less attractive firefly i
et al., 2013). The farther another firefly, the less bright it will towards a more attractiveness firefly j is given by (Yang,
be to the vision of the first firefly. This is due to the light 2009)
intensity decreasing under the inverse square law. That is, the
light intensity of a firefly with ‘‘r’’ distance between any two 2 1
xi = xi + b0 egrij (xj  xi ) + a(rand  ) ð10Þ
fireflies will be reduced by a factor of 1=r. The air will also 2
absorb part of the light as the firefly travels, which will further
where xi is the current position of the firefly within the solu-
reduce the perceived intensity. The light intensity I(r) varies
tion space. The combination of the elements in the second
according to the following equation (Rahkar-Farshi and
term represents the firefly’s attractiveness, as seen by the other
Behjat-Jamal, 2016)
fireflies. A random adjustment is given in the movement of
2 the firefly with a representing a scaling parameter to control
I(r) = I0 egr ð7Þ the amount of randomness. The scaling parameter a 2 ½0, 1,
with being
where I0 denotes the light intensity of the source, g 2 ½0, ‘
represents the light absorption coefficient that controls the 
0 Corresponds to no randomness
decrease in the intensity of light. ð11Þ
1 Corresponds to fully being random
The attractiveness, b, of fireflies is proportional to their
light intensities, I(r). Therefore, a similar equation to (7) can
Notice that rand (a MATLAB Command) is a uniformly dis-
be defined to describe the attractiveness b (Rahkar-Farshi
tributed random number, in which rand 2 (0, 1) (Yang, 2009).
and Behjat-Jamal, 2016)
2
b(r) = b0 egr ð8Þ FA optimization approach
where b0 is the initial attractiveness at (r = 0). FA is based on the objective function and corresponding fit-
To start the search operation algorithm, the fireflies are ness levels used to find the optimal solution. It is an iterative
placed in random locations. The location of a firefly corre-
sponds to the values of the parameters for the objective func-
tion to be solved (Rahkar-Farshi and Behjat-Jamal, 2016).
For any two flashing fireflies, as seen in Figure (1), the less
bright firefly (FFi) moves toward the brighter one (FFj)
according to the attractiveness b, regardless of their sex. The
attractiveness is proportional to the light intensity, I, observed
by the partner, which monotonically decreases as the distance
between two fireflies increases. This is due to the inverse
square law and the absorption coefficient of light. The bright-
est firefly moves randomly (Hachino et al., 2015).
The FA consists of only the basic arithmetic operations
and does not require complicated coding and genetic opera-
tions such as crossovers and mutations of the GA. In addi-
tion, the performance and computational cost of the FA are
shown to be better than those of other population-based algo-
rithms such as the GA and the POS (Lohrer, 2013). These Figure 1. Fireflies formatting search process.
2878 Transactions of the Institute of Measurement and Control 41(10)

optimization procedure as it works with a population that x_ r ðtÞ = Ar xr ðtÞ + Br uðtÞ,


represents a number of solutions rather than a single solution ð12Þ
yr ðtÞ = Cr xr ðtÞ + Dr uðtÞ
in each iteration. It makes the decision of its solutions by eval-
uating some fitness function. The final solution is obtained by Depending on the order to be reduced, the poles nearest to
updating that fitness function corresponding to some given the origin (dominant) are retained. This implies that the over-
specifications. FA can be controlled by three parameters: the all behavior of the reduced system will be very similar to the
randomization parameter a, the attractiveness b, and the original system, since the contribution of the unretained
absorption coefficient g. According to the parameter setting, eigenvalues to the system response are important only at the
FA distinguishes two asymptotic behaviors. The former beginning of the response, whereas the eigenvalues retained
appears when g = 0 and the latter when g = ‘. If g = 0, the are important throughout the whole of the response.
attractiveness becomes b = bo . That is, the attractiveness is Accordingly, the FA would then determine the rest of the
constant anywhere within the search space. This behavior is a reduced-order model elements seen in the Br and Cr matrices
special case of PSO. If g = ‘, the second term falls out from given as
equation (10), and the firefly movement becomes a random 2 3
walk, which is essentially a parallel version of simulated b11  b1m
annealing. In fact, each implementation of FA can be between 6 . .. .. 7
Br = 4 .. . . 5 ð13Þ
these two asymptotic behaviors (Fister et al., 2013). bnr 1  bnr m
To obtain the reduced-order models, the light intensity (fit-
ness value) of the firefly uses the inverse of the integral of the 2 3
c11  c1nr
magnitude squared of the frequency-weighted model error 6 .. .. .. 7
Cr = 4 . . . 5 ð14Þ
between the original system and the reduced order model. The
best arguments of the model reduction are achieved through bp1  cpnr
searching by the fireflies, and are considered as the result of
As a result, one can see that the number of parameters to be
the last search. The firefly MOR procedure is illustrated in
estimated by the FA is
Figure 2.
Applying the FA optimization technique for MOR, the
np = nr  m + p  nr ð15Þ
following steps were taken:
It is important to mention that for equations (5) and (6),
(1) Determine the dominant eigenvalues that will keep the reduction is performed based on the fact that the system
the behavior of the reduced system closer to the origi- is a multi-time-scale type. That is, there exist two categories
nal system. This step will produce Ar . in the system, slow and fast, which are distinguished by a fac-
(2) Using FA, Br and Cr will be determined. The follow- tor of 10 for a proper MOR as motivated by the SPA method
ing state-space representation is obtained (Alsmadi et al., 2014a).
To compare the results of the different algorithms, the root
mean square error (RMSE) was used
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N
u
2 1
X
RMSE = t ðy  yr Þ2 ð16Þ
N i=1

where N is the size of the input time.

Simulation examples
This section presents some simulation results and comparison
diagrams that have been used to compare the proposed FA
approach with other known algorithms. MATLAB software
tools were utilized for the simulations. The parameters of the
proposed FA technique were used to design the state matrix
are shown in Table 1.

Example 1
As an illustrative example, we will consider a Single-Machine
Infinite-Bus (SMIB) power system, shown in Figure 3, to
examine the proposed FA approach. The parameter values
are given as specified in Parmar et al. (2007). The machine is
supplying power through a step-up transformer and a high-
Figure 2. FA MOR flowchart. voltage transmission line to an infinite grid. The XT and XL in
Alsmadi et al. 2879

Table 1. FA parameters.

Parameter Value

Firefly size (M) 50


Attractiveness (b) 1
Media absorption coefficient (g) 1
Randomness reduction (a) .98
Randomness (d) 1
Maximum Number of iteration 500
Initial minimum value of the cost function (minVal) 13106
Upper limit (upperlim) 5
Lower limit (lowerlim) 25

Figure 4. System eigenvalue constellation.

closer poles to the origin. As can be seen in the set of eigenva-


lues, the multi-time-scale system can be categorized as a
Figure 3. Power system for a SMIB. two-time-scale system (slow and fast subsystems). The fast-
subsystem is given with eigenvalues lf = {-18.9311 6 2.0250i,
-12.1968, -9.6484}, while the slow-subsystem eigenvalues may
be given as: ls = {-2.1313, -0.8972 6 1.3560i, -0.2394 6
the figure represent the reactance of the transformer and the
3.2350i,-0.1001}. Hence, the proposed reduction is performed
transmission line, respectively. The VT and VB are the genera-
tor terminal and infinite bus voltages, respectively. by selecting the slow-subsystem and eliminating the fast sub-
The system consists of a three-phase 160-MVA synchro- system, as illustrated by Gharaibeh (2016) and as shown in
nous machine with automatic excitation control system. In a Figure 4. Thus, the 6th order reduced model is investigated.
state space representation, the system is given as follows
2 3
0:5517 0 0:3091 0 0 0 0 0 0 0:1695
6 0:0410 0 0:0350 0 0 0 0 0 7 0 0
6 7
6 7
6 0 314:1593 0 0 0 0 0 0 7 0 0
6 7
6 9:5540 0 0:8660 20 0 0 0 0 0 0 7
6 7
6 7
6 0 0 0 0 1 0 0 0 0:0421 0:0328 7
Af = 6
6 0:1962
7
7
6 10:8696 0:1672 0 0 10:8696 0 0 0 0 7
6 7
6 0:9386 51:9849 0:7999 0 0 41:1153 10:8696 0 0 0 7
6 7
6 0:9386 51:9849 0:7999 0 0 41:1153 10:86:96 0:1 0 0 7
6 7
6 7
4 0 0 0 1000 1000 0 0 1000 20 0 5
0 0 0 0 0 0 0 0 1:0526 0:8211
 
0 0 0 0 0 0 0 0 1000 0 T
Bf =
0 0:0926 0 0 0 0:4428 2:1179 2:1179 0 0
   
0 0 1 0 0 0 0 0 0 0 0 0
Cf = , Df =
0:4777 0 0:0433 0 0 0 0 0 0 0 0 0

with eigenvalues given as: l = {-18.9311 6 2.0250i, -12.1968, The proposed method is first investigated for SISO type
-9.6484, -2.1313, -0.8972 6 1.3560i, -0.2394 6 3.2350i,- systems, which for this example, the second column of the sys-
0.1001}. tem input matrix and the second row of the output system
Indeed, there exist different methods for determining the matrix are eliminated. Thus, performing the reduction opera-
dominant dynamics of a given system. In this paper, however, tion, the following 6th order reduced model was obtained
we will use the traditional method, which is by inspecting the
2880 Transactions of the Institute of Measurement and Control 41(10)

Figure 7. Frequency responses for the 6th reduced order and the 10th
Figure 5. Step responses for the 6th reduced order and the 10th full
full order models.
order models.

model response. Investigating the robustness of the proposed


MOR approach, the following 5th order model is obtained
2 3
0:2394 3:2350 0 0 0
6 7
6 3:2350 0:2394 0 0 0 7
6 7
x_ r ðtÞ = 6
6 0 0 0:8972 1:3560 0 7
7
6 7
4 0 0 1:3560 0:8972 0 5
0 0 0 0 0:1001
2 3
1:5263
6 7
6 1:2728 7
6 7
xr ð t Þ + 6 7
6 1:5979 7u(t)
6 7
4 0:6947: 5
0:0089
yr ðtÞ = ½0:1267  2:1455  3:0861  1:02771:8082 xr ðtÞ

The selection of this 5th order model shows that the dominant
Figure 6. Impulse responses for the 6th reduced order and the 10th full dynamics of the original system presented exactly in the
order models. reduced order model, which can be seen as

2 3 2 3
0:2394 3:2350 0 0 0 0 1:8080
6 3:2350 0:2394 0 0 0 0 7 6 2:2090 7
6 7 6 7
6 7 6 7
6 0 0 0:8972 1:3560 0 0 7 6 1:2650 7
x_ r ðtÞ = 6
6
7xr ðtÞ + 6
7
7
6 0:8226 7uðtÞ
6 0 0 1:3560 0:8972 0 0 7 6 7
6 7 6 7
4 0 0 0 0 2:1313 0 5 4 0:7837 5
0 0 0 0 0 0:1001 0:07042
yr ðtÞ = ½1:324 0:1057  0:9861  4:997 4:026  0:01518 xr ðtÞ

with step and impulse responses, along with the frequency


response, presented as shown in Figures (5), (6), and (7), l = f0:2394 6 3:2350i,  0:8972 6 1:3560i,  0:1001g
respectively, compared with the original 10th order models’.
It can be seen that the performance of the proposed The step responses of this reduced 5th order model, along
reduced order FA model is almost identical to the original with the original full 10th order models, are shown in
Figure 8.
Alsmadi et al. 2881

Figure 9. System cost minimization convergence for a step input


Figure 8. Step responses for the 5th reduced order (dark solid line)
response.
and the 10th full order models.

For a rigorous investigation of the proposed method, we


now consider the original MIMO system with farther reduc-
tion performed to a 4th order model. This means that we will
eliminate two of the dominant system dynamics. As can be
seen from Figure 4, they would have to be {20.8972 6
1.3560i}, which makes the slow category then be given by ls
= {22.1313, 20.2394 6 3.2350i, 20.1001}. However, due to
the relatively high natural frequency seen in the dominant
poles {20.8972 6 1.3560i}, we will instead eliminate the
dynamics {22.1313, 20.1001}. Thus, the reduced model will
have the slow category given by ls = {20.8972 6 1.3560i
20.2394 6 3.2350i}. As a result, using the proposed FA-
MOR, the 10th order model was reduced to a 4th order with
system elements given by
2 3
0:2394 3:2350 0 0
6 3:2350 0:2394 0 0 7
6 7
Ar = 6 7, Figure 10. Step responses for the 4th reduced order (dotted line) and
4 0 0 0:8972 1:3560 5
the 10th full order models.
0 0 1:3560 0:8972
2 3
0:0050 1:7083
6 0:9596 1:6849 7 To investigate the performance of the FA-MOR method,
6 7
Br = 6 7
4 0:4723 0:4902 5 the full and reduced order models were both simulated to step
1:8536 0:3243 inputs with responses for both outputs presented as shown in
  Figure 10.
3:3957 2:6668 2:3774 1:5262
Cr = , As seen in Figure 10, the reduced 4th order model
0:2601 0:0392 0:8995 0:1644 responses are relatively very close to the full 10th order mod-
 
0 0 els’ seen in both stages, the transient and steady state. When
Dr =
0 0 comparing the performance of the FA-MOR with the firefly
technique proposed in Alsmadi et al. (2016), the responses of
Notice that the substructure preservation has been achieved the two methods along with the original system responses are
as seen in the Ar matrix, as it contains the exact dominant presented as shown in Figure 11. Alsmadi et al. (2016) con-
eigenvalues of the full order model. That is, the dominant cluded that the FA is not suitable for MOR based on their
dynamics of the reduced order model are given as {20.8972 observations and method parameter design. As can be clearly
6 1.3560i, 20.2394 6 3.2350i}, which are basically a subset seen now, the proposed FA method provides much better
of the full 10th order model. The operation was performed to responses, in which the design of the parameters were found
a step input with a fitness progress presented by ‘cost value’, to be best at beta = 1, gama = 1, alpha = 0.98, and zeta =
as shown in Figure 9. 1. Hence, it can be concluded that the FA-MOR proposed
2882 Transactions of the Institute of Measurement and Control 41(10)

with eigenvalues given as {-1, -2, -3, -4, -5, -6, -7, -8}.
Observing the dynamics (eigenvalues) of this system, it is seen
that the system is not a two-time scale type; however, it is
found that the dominant dynamics are seen in the poles {-1, -
6}. Hence, the reduction process is performed correspond-
ingly. As a result, the following 2nd order reduced model is
obtained

17:198052s + 6:0524
Gr (s) =
s2 + 7s + 6

Desai and Prasad (2013) and Abu-Al-Nadi et al. (2011) also


found, correspondingly, the following 2nd reduced orders,

16:91s + 5:255
Gmodel1 (s) =
s2
+ 6:87s + 5:26
17:0989s + 5:0742
Gmodel2 (s) = 2
Figure 11. Step responses for the full 10th order (solid line), firefly s + 6:9722s + 5:1514
reduced 4th order (dotted light line), and proposed FA reduced 4th
order models (dotted dark line).
The simulation results are presented in Figure 12. In part of
the comparison, it is important to mention that the original
exact dominant dynamics are only preserved by the new
approach.
method is superior and has a potential advantage in provid-
ing reduced order models with SP and relative accuracy of
system response. Example 3
In this example, we consider a 9th order system evaluated by
Example 2 Desai and Prasad (2013) and Boby and Pal (2010). The system
is given by the following transfer function
In this example, we consider an 8th order SISO system evalu-
ated by Desai and Prasad (2013) and Mukherjee et al. (2005).
The system is given as follows

18s7 + 514s6 + 5982s5 + 36380s4 + 122664s3 + 222088s2 + 185760s + 40320


G(s) =
s8 + 36s7 + 546s6 + 4536s5 + 22449s4 + 67284s3 + 118124s2 + 109584s + 40320

Figure 12. Step and frequency responses for the full 8th order, Model1 reduced 2nd order, Model2 reduced 2nd order and proposed FA reduced 2nd
order models.
Alsmadi et al. 2883

Figure 13. Step and frequency responses for the full 8th order, Model1 reduced 2nd order, Model2 reduced 2nd order and proposed FA reduced 2nd
order models.

s4 + 35s3 + 291s2 + 1091s + 1700


G(s) =
s9 + 9s8 + 66s7 + 546s6 + 294s5 + 2541s4 + 6484s3 + 5856s2 + 4620s + 1700

with eigenvalues given as {-1, -161i, -162i, -163i, -164i}. methods is based on the values of the RMSE, which repre-
Observing the dynamics (eigenvalues) of this system, it is seen sented as equation (16).
that the system is not a two-time scale type; however, it is con- The results in Table (2) show that proposed FA has the
sidered for comparison purposes, keeping in mind that preser- second-best solution after PSO optimization with RMSE of
ving the exact dominant dynamics in the reduced order is only 0.002.
performed in the proposed method. Table (3) presents the evaluations of FA results in Alsmadi
Performing the FA reduction technique, and focusing on et al. (2016) and the proposed FA, with results seen as much
the dynamics {-1, -161i}, the following 3rd order reduced better in the proposed work.
model is obtained When FA simulation results were compared with other
methods’ results, such as GA, PSO algorithm and IWO, in
0:50532s2 + 0:4003s + 2:0 solving MOR problems, it was found that FA transcends the
Gr (s) =
s3 + 3s2 + 4s + 2 other methods in the error sense by leading to better (lower)
RMSE. Also, it is noted that the FA was more accurate and
while Desai and Prasad (2013) and Boby and Pal (2010), cor- the success rate was remarkably better than the other men-
respondingly, found the following reduced orders, tioned algorithms.
0:08717s2 + 0:3142s + 0:493
Gmodel1 (s) =
s3 + 1:494s2 + 1:34s + 0:493
0:5058s2  1:9848s + 3:5341
Gmodel2 (s) = 3
s + 3s2 + 5:5341s + 3:5341
which provided the simulation results shown in Figure 13. In
short, it is seen that each method has its advantages and
shortcomings.

Discussion
In comparing the performance of the proposed method with
others, Figure (14) shows five different diagrams, comparing
between the original system of example one, without reduc-
tion and reduced models using four different optimization
techniques: IWO (Mehrabian and Lucas, 2006), PSO
(Kennedy and Eberhart, 1995), GA (Abo-Hammour et al.,
2011) and the proposed algorithm (firefly optimization) for
fourth order reduction. The comparison between these Figure 14. Comparison of step responses for four different techniques.
2884 Transactions of the Institute of Measurement and Control 41(10)

Table 2. Evaluation of methods. Alsmadi OMK, Abo-Hammour ZS, Al-Smadi AM and Abu-Al-
Nadi DI (2011b) Genetic algorithm approach with frequency
Method Generation RMSE SP selectivity for model order reduction of MIMO systems. Mathe-
matical and Computer Modelling of Dynamical Systems 17(2):
GA 200 0.0033 Achieved 163–181.
IWO 1000 0.0038 Achieved Alsmadi OMK, Abo-Hammour ZS, Al-Smadi AM and Saraireh SS
PSO 500 0.0024 Achieved (2011c) A robust and efficient genetic algorithm for solving a
Proposed FA 500 0.0025 Achieved chemical reactor problem: Theory, application and convergence.
Transactions of the Institute of Measurement and Control 34(5):
594–603.
Alsmadi OMK, Abo-Hammour ZS, Saraireh SS and Marzouq AA
(2014a) Substructure preservation sylvester-based model order
Table 3. Evaluation of methods.
reduction with application to power systems. Electric Power Com-
ponents and Systems 42(9): 914–926.
Method Generation RMS SP Steady State
Alsmadi OMK, Abu-Al-Nadi DI and Abo-Hammour ZS (2014b)
error
Partical swarm optimaization for MOR singularly perturbed sys-
FA paper of 2500 0.0289 Achieved Not Achieved tems with critical frequency preservation and application to power
Alsmadi et al. systems simplified modeling. Journal of Circuits, Systems and
(2016) Computers 23(5): 1–20.
Proposed FA 500 0.0020 Achieved Achieved Antoulas AC, Sorensen DC and Gugercin S (2001) A survey of model
reduction methods for large-scale systems. Contemporary mathe-
matics 280: 193–220.
Benner P, Gugercin S and Willcox K (2015) A survey of projection-
based model reduction methods for parametric dynamical systems.
Conclusions SIAM Review 57(4): 483–531.
We examined the problem of MOR for multi-time scale sys- Boby P and Pal J (2010) An evolutionary computation based
tems. The reduction process was performed using the firefly approach for reduced order modelling of linear systems. In: IEEE
International Conference on Computational Intelligence and Com-
optimization technique. This approach was able to produce
puting Research (ICCIC), Coimbatore, India, 28–29 December
reduced order models with the advantage of substructure pre- 2010.
servation, as to maintaining the system dominant dynamics in Chen C and Shieh L (1968) A Novel Approach to Linear Model Sim-
the reduced order. The performance of the proposed method plification. International Journal on Control 8(2): 561–570.
was compared with recently published work concerning MOR Chidambara MR (1967) Further Comments by M.R. Chidambara.
via the firefly method. Simulation results show the potential IEEE Transaction on Automatic Control AC-12: 799–800.
of the FA as an artificial intelligence technique for the process Davison EJ (1966) A Method for Simplifying Linear Dynamic Sys-
of MOR. In addition, the results of the proposed FA reduc- tems. IEEE Transactions on Automatic Control AC-11(1): 93–101.
tion technique were compared with three other optimization Desai SR and Prasad R (2013) A new approach to order reductionus-
ing stabilityequation and big bang big crunch optimization. Sys-
techniques, namely, IWO, PSO and GA. The results show
tems Science and Control Engineering Journal 1(1): 20–27.
some superiority of the proposed FA over the other methods.
Fister I, Yang X and Brest J (2013) A comprehensive review of firefly
algorithms. Swarm and Evolutionary Computation 13(3): 34–46.
Funding Gharaibeh EA (2016) Design of reduced order controller for multi-
time-scale power systems via firefly optimization. MS Thesis, Yar-
The author(s) received no financial support for the research, mouk University, Irbid, Jordan.
authorship, and/or publication of this article. Gugercin S and Antoulas AC (2004) A survey of model reduction by
balanced truncation and some new results. International Journal of
Control 77(8): 748–766.
References Hachino T, Tanigawa K, Takata H, Fukushima S and Igarashi Y
Abo-Hammour ZS, Asasfeh AG, Al-Smadi AM and Alsmadi OMK (2015) Frequency-weighted model reduction using firefly algo-
(2011) A novel continuous genetic algorithm for the solution of rithm. Journal of Automation and Control Engineering 3(3):
optimal control problems. Optimal Control Applications and Meth- 222–227.
ods 32(4): 414–432. Iglesias A (2013) Firefly Algorithm for explicit b-spline curve fitting
Abu-Al-Nadi DI, Alsmadi OMK and Abo-Hammour ZS (2011) to data points. Mathematical Problems in Engineering 2013: 1–12.
Reduced order modeling of linear MIMO systems using particle Kennedy J and Eberhart R (1995) Particle swarm optimization. In:
swarm optimization. The Seventh International Conference on IEEE International Conference on Neural Networks 4, pp. 1942–
Autonomic and Autonomous Systems, Venice, Italy, May 2011, pp. 1948. Perth, WA: IEEE.
62–66. IARIA. Lohrer MF (2013) A comparison between the firefly algorithm and
Alsmadi OMK, Abo-Hammour ZS, Abu-Al-Nadi DI and Saraireh particle swarm. Dissertation, Graduate Honor College, Oakland
SS (2016) Soft Computing Techniques for Reduced Order Model- University.
ing: Review and Application. Intelligent Automation & Soft Com- Mehrabian AR and Lucas C (2006) A novel numerical optimization
puting 22(1): 125–142. algorithm inspired from weed colonization. Ecological informatics
Alsmadi OMK, Abo-Hammour ZS and Al-Smadi AM (2011a) Artifi- 1: 355–366.
tial neural network for the discrete model order reduction with Mukherjee S, Satakshi S and Mittal RC (2005) Model order reduction
substructure preservation. Applied Mathematical Modelling 35(9): using response-matching technique. Journal of the Franklin Insti-
4620–4629. tute 342(5): 503–519.
Alsmadi et al. 2885

Parmar G, Mukherjee S and Prasad R (2007) Reduced order model- Yadav JS, Patidar NP, Panda S and Ardil C (2012) evolutionary tech-
ling of linear MIMO systems using genetic algorithm. Interna- niques for model order reduction of large scale linear systems.
tional Journal of Simulation Modelling 6(3): 173–184. International Journal of Electrical, Computer, Electronics and Com-
Rahkar-Farshi T and Behjat-Jamal S (2016) A multimodal firefly opti- munication Engineering 6(9): 1–7.
mization algorithm based on Coulomb’s Law. International Journal Yang XS (2009) Firefly algorithms for multimodal optimization. Sto-
of Advanced Computer Science and Applications 7(5): 134–141. chastic Algorithms: Foundations and Applications, [5th Interna-
Salimbahrami B and Lohmann B (2006) Order reduction of large tional Symposium [Lecture Notes in Computer Science book series
scale second-order systems using Krylov subspace methods. Lin- (LNCS, Volume 5792)] Sappora, Jappan, 26–28 October 2009,
ear Algebra and its Applications 415: 385–405. pp. 169–178.
Wilson DA (1970) Optimum solution of model-reduction problem.
Proceedings of the Institution of Electrical Engineering. IEEE
117(6): 1161–1165.

You might also like